jiwo
Small open decision models (0.8B and 4B, from Qwen3.5) with a server that speaks Jev's /v1/systemone format in one forward pass.
Overview
jiwo provides small open decision models that read a state and typed questions and return a probability for every option in one forward pass, without generating text. They are fine-tuned from Qwen3.5-0.8B and Qwen3.5-4B under Apache-2.0, and served with the same request and response format as TypeSafe's Jev API, POST /v1/systemone. It is independent and not endorsed by TypeSafe.
Both models are on the Jev Decision Index board. The README reports v1 of jiwo-0.8b ranked first of 27 models under 1B and jiwo-4b third of 26 in the 3 to 6B class, with v2 public scores of 29.58 and 46.06, well above the untrained base models under the same server and calibration. Median latency is about 50 ms on an H100, and 13.1 ms or 31.6 ms on an RTX PRO 6000 with CUDA graphs. Training data is mostly English, and other languages are not evaluated. It installs from GitHub with Python 3.12 and runs on GPU or CPU (float32 recommended on CPU).
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Any yes/no question about the state (e.g. The customer is angry.) | — |
| Choice | Which option fits the state? | — |
| Score | Where does the state fall on the scale? | — |
How it works
- Install jiwo and load a model from Hugging Face
- Send state and typed questions via Python or the HTTP server
- Model returns a probability per option in one pass
- Your code thresholds or branches on the probabilities
Strengths
- Open weights under Apache-2.0
- Jev-compatible API
- Very low latency on modern GPUs
- Ranked on an independent leaderboard
Caveats
- Mostly English training data
- Full Decision Index scores for v2 were pending
- Smaller models trail hosted Jev on hard tasks
Getting started
pip install "jiwo @ git+https://github.com/jiwidi/jiwo", then DecisionModel.from_pretrained("eljiwo/jiwo-0.8b").At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- PyTorch library and HTTP server
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Small open decision models fine-tuned from Qwen3.5 (0.8B and 4B, Apache-2.0) with a server that speaks Jev's `POST /v1/systemone` format: one forward pass, a probability per option.
MrJev/awesome-jev
Frequently asked questions
Can I use a Jev client with jiwo?
Yes, the server uses Jev's /v1/systemone request and response format.
How good are the models?
The README cites v1 ranking first under 1B and third in the 3-6B class on the Decision Index.
Does it work on CPU?
Yes, using float32; bfloat16 is slow on many CPUs and Apple silicon.
Related projects
Toy local Jev-shaped decision API on a 0.5B model that answers yes/no, choice and score questions by scoring option tokens.
ServiceprototypePythonServes MiniCPM5-2B-8bit on Apple Silicon as a Jev-compatible /v1/systemone service and reproduces its JevBench score.
ModelprototypePython- reflex★ 166
Small open decision model on a frozen Qwen3.5-4B serving /v1/systemone in about 200 ms, benchmarked against Jev on JevBench.
ServiceresearchPython Open Apache-2.0 Jev-like model family on Qwen3.8-27B that reads one-token candidate preferences and serves a TypeSafe-compatible API.
ServiceresearchPythonOpen MiniCPM5-2B decision model serving /v1/systemone: typed choice, noul and score probabilities in one forward pass.
ModelresearchPythonDocker recipes and ARM64 images for serving ten native decision checkpoints, such as Cloudflare Clef, behind /v1/systemone on DGX Spark.
ServicebetaPython